Some time ago I heard a song named ‘a Fifth of Beethoven’ by Walter Murphy and I really liked it. It is a disco instrumental song based on Beethoven’s Symphony No. 5. I thought using famous classical music pieces to create new songs of another genre is a really cool concept. So I did some research and discovered that there are many more famous songs that are based and inspired on a classical piece of music, also songs I know for a long time but never knew that they have a classical origin.
My corpus includes classical pieces and their linked new songs.
I have found 25 pairs, so 50 songs in total.
I think that this corpus makes really interesting comparison possible, on different levels.
For example between the songs based on the classical songs.
What are the differences and similarities between these songs?
Is the classical input used in the same way? Is there an common genre?
But of course exploring relation between the song based on a classic song and the original classical song itself is really interesting.
To what extent do you see the original classical back in the new songs? In which way is the original song adapted in de new song? Which changes are made? Maybe the tempo changed or other instruments are used.
There are pairs of tracks in the corpus, like ‘I can’ – Nas/‘Für Elise’ -Beethoven and ‘A fifth of Beethoven’-Woody/ Beethoven’s Fifth Symphony- Beethoven, in which the relation with the origin classical song is very obvious. But there are also pairs in the corpus in which the original songs are more subtly included, like ‘Grace Kelly’- Mika / ‘Largo Al Factotum’- Rossini.
The corpus isn’t perfect. It would be better if the playlist of the inspired newer songs would also inlcude only one genre instead of all different kinds. I that was the case later conclusions would be more reliable than now, because the likelihood that other factors play a role would be smaller. It’s good to be aware of that while reading portfolio. Nevertheless I think it’s an interesting corpus to do research on.
>
| New | |
|---|---|
| I can - Nas | Fur Elise - Beethoven |
| A fifth of Beethoven - Walter Murphy | Symphony No, 5 in C minor - Beethoven |
| Because - The Beatles | Sonata No. 14 “Moonlight” in C sharp - Beethoven |
| Can’t Help falling in Love - Elvis Presley | Martini: Plaisir d’amour - Jean-Paul-Egide Martini |
| All by myself - Eric Carmen | Piano Concerto No. 2 in C minor - Sergei Rachmaninoff |
| The Globalist - Muse | Nimord From Enigma Variations - Edward Elgar |
| Say You’ll Go | Clair de Lune, L.32 - Claude Debussy |
| Old money - Lana Del Rey | A time for Us - Nino Rota, Angele Dubeau |
| Little Me - Little Mix | Pavane, Op. 50 - Gabriel Faure |
| I Believe in Father Christmas - Greg Lake | Prokofiev: suite from lieutenant kijé - Sergei Prokofiev |
| Road to Joy - Bright Eyes | Symphony No. 9 in D Minor, Op. 125 - Ludwig van Beethoven |
| It’s A Hard Life - Queen | Vesti la giubba - Ruggero Leoncavallo |
| Tocacata - Sky | Toccata and Fugue in D minor - Johann Sebastian Bach |
| A Whiter Shade of Pale - Procol Harum | Air on a G string - Johann Sebastian |
| Could it Be Magic - Barry Manilow | Preludes, Op. 28: No 20 in C minor - Frederic Chopin |
| Go West - Pet Shop Boys | Canon in D - Johann Pachelbel |
| Alejandro - Lady Gaga | Csardas - Vittorio Monti |
| Ave Maria - Beyonce | Schubert: Ave Maria, D. 839 - Franz Schubert |
| Sara - Starship | Fantasia on a Theme by Thomas Tallis - Ralph Vaughan Williams |
| Grace Kelly - MIKA | Largo al factotum - Gioachino Rossini |
| Exit Music (For a Film) - Radiohead | Prelude in E minor, Op. 28, No. 4 - Frederic Chopin |
| Bad Romance - Lady Gaga | The Well-Tempered Clavier: Book 1 - Johann Sebastian Bach |
| Russians - Sting | Lieutenant Kijé, Opt. 60: 2. Romance - Sergei Prokofiev |
| My Reverie - Ella Fitzgerald | Reverie - Claude Debussy |
| Hey Jude - The Beatles | Arioso (Adagio in G) - Johann Sebatian Bach |
In this graph you can have a look at the different features of the inspired songs and the features of the original songs. The lines in the plot show if there is a decrease or increase in value of the different features. A yellow-ish line means that from the inspired newer song to the original classical song, there is an increase of the feature. A blue-ish line means that from inspired newer song to the original classical song, there is an decrease of the feature.
difference of the features (highest differences on top) inspired newer songs -> original classical songs - instrumentalness: 0.527 - acousticness: 0.521 - energy: -0.428 - valence: -0.210 - danceability: -0.170 - speechiness: 0.129 - liveness: -0.039
This plot shows 4 variables. The x-axis variance, the y-axis energy, the size of the dots is loudness and the color is the mode.It points out some difference between the classical songs and the newer songs. It’s interesting to see that the the values of the features of the classical songs are more clustered, while the ones of newer songs are more spread out. A possible explanation could be the fact that the original classical songs all are all in the same genre; classical music. The playlist of the inspired newer songs, on the other hand, contains more genres, like pop, hiphop and jazz.
Because of the different scale, loudness isn’t showed in the overview graph on the previous page, but it’s an interesting feature, so I used a different graph to show this. The histogram of the loudness feature shows a clear distinction in values between the two playlists. The newer songs are a lot louder.
In the graphs on the left you see the self-similarity matrices of Air on a G string by Bach and Fur Elise by Beethoven. Both Chroma-features as Timbre-Features.
In these SSM it’s visible that:
The paths parallel to the diagonal shows that there is in both song repetition, especially in fur Elise. If you listen to Fur Elise you can notice many repetition parts.
The self-similarity matrices of Air on a G string is less clear
The blocks in Fur Elise show that the clear homogeneity Fur Elise.
The point where the corners of the blocks touch is the part where there appears a turning point in the song. This is calles novelty
I compared two song of the corpus using key-graphs. The first graph is a visualization of ‘My Reverie’ by Ella Fitzgerald. This song contains some clear repetitions. The second graph is a visualization of the classical song that inspired Ella Fitzgerald, named ‘Rêverie’ by Claude Debussy.
An interesting observation is that they have some clear similarities. Although Debussy’s Rêverie is bit more stretched out, you can spot the same pattern. This is also something that you can hear back if you listen to the two songs. The lighter parts in the middle and in the end are in both pieces audible. Both of the songs sound louder, but in a different way. In Fitzgerald’s song increases the sound other voice and in Debussy’s song increases the sound of the instruments.
This chromagram is made using the Dynamic Time-Warping technique (DTW), using Manhattan normalization and Aitchison distance to align both versions. DTW is a great help if you want to compare various performances. You can see by the longer x-axis that the Bach Collegium Japan plays much slower, resulting in a rectangular chromagram.
My expectations were high for the DTW because the Dynamic Time-Warping technique (DTW) is a good tool to compare different performances. In a way, you could see the inspired newer songs as a different performance of the original song, so this gave high hopes.
Unfortunately it didn’t turn out to work as good as expected. After trying all the possible normalization and distances on different songs of the corpus, I could conclude that this technique wasn’t working for my corpus.
So I started thinking of reasons why it failed.
As you can see in the histogram, the tempi of the classical songs vary a lot. It’s almost an evenly distributed graph. There are only a few more slower songs than faster songs. The different tempi of the inspired newer songs are closer to each other, less distributed. There is a peak for the newer songs around 120 bmp.
This is interesting because you could expect it the other way around. The playlist of the original classical songs is only one genre, while the playlist of the inspired newer songs contains songs of many different genres. First I though this could affect the distribution of the tempi of the songs, in a way that the playlist with the inspired newer songs would be spread out a lot, but it turns out to be the opposite. Most of the inspired newer songs got a tempo around 120BMP, while the distribution of the classical songs doesn’t show a peak in the graph. Apparently classical music is very versatile in the field of tempo!
On this page you can find 4 chromagrams. It is interesting to see that some original classical songs and its newer inspired version show really clear similarities while other duo’s don’t. Some duo’s are played in the same pitch, while others are a few semitones lower or higher.
An example of similarities are the first two chomagrams: Sonata No. 14 “Moonlight” in C sharp - Beethoven and the inspired Because - The Beatles . It’s played in C sharp (surprise! ;)), G sharp and B.
An example of differences in pitches are the third and fourth chromagrams: Prelude in E minor, Op. 28, No. 4 - Frederic Chopin and the inspired Exit Music (For a Film) - Radiohead . Chopin plays the song in E minor, while Radiohead plays it in F sharp.
This isn’t really weird, because it’s possible to be inspired by a song, but change the pitches as well. But it;s interesting to see that some inspired songs also are influenced by the pitch, while other inspired songs change the pitch
.
First we optained the importance for each feauture, by training a random forest classifier. The loudness and acousticness turned out to be the best features to use. (this is in line with the largest differences between the features calculated in the first pages of the portfolio)
After classifying the two playlist we got an overall precision of 71%. So it did quiet a good job!
Of course I learned a lot about the Spotify API and all it’s possible techniques, but what In previous courses of my study I learned extract information of large data sets . I did use these quantative research skills a lot and they were really useful for my research, but during this course I also learned to zoom in on specific datapoints (in this case individual songs) and apply more qualitatively research.
So two levels of comparisments are made in this portfolio.
The main findings:
The differences and similarities between the classical song and its inspired song, really differ per duo. There is no spotify API technique that shows for all duo’s similarities.
Eventhough you the newer songs are inspired by the classical songs, there is too much deviation to find a warping path
It’s often hard to find the similarities between the classical song and its inspired newer song.
You could see the large varity of duo’s as a limitation, but the he fact that we can conclude that finding multiple similarities isn’t that obvious and that there isn’t one spotify API tool that works for all duo’s similarities, also says something interesting about the corpus that I discovered during this course: There are multible ways to incorporate a classical song into a newer inspired version, and these different ways can be shown by using different Spotify API techniques, like chordograms, histograms, self-similarity matrices and key-graphs.
When comparing the playlist with classical songs with the playlist with new songs, I thought I mainly would find obvious differences, such as feature differences that appear if you compare classical songs with newer songs. But after applying a some spotify API techniques, I also started to notice that there is a big overall difference between the two playlists. The songs of the classical playlist appear to differ a lot from each other, compared to the newer songs that had more similarities. This was visible in the graphs. The values in the graphs of the classical songs were a lot more spread out. The newer songs, on the other hand, showed clustering and peaks. I though this is remarkable, because you could expect the classical songs to show more similarities, given that songs belong to the same genre.
I really enjoyed making this portfolio. In the beginning I had some doubts if my corpus was well fitted for using the Spotify API. I had the idea that the Spotify API was mostly mend to be used on large scale, like multiple playlists with many songs. While the interesting part of this corpus, is more zoomed in, namely comparing the original classical song with its inspired newer version. But during the course I found out that Spotify API contain all kind of functions, also many individually focussed techniques, like the chromagrams, key-graphs and DTW. These tools were perfect for comparing two songs. The spotify API turned out to be super versatile and I learned using many of the techniques it offers.